"""Template for showing the results of the last experiment in MLFlow."""

import logging
import numpy as np
import helpsk as hlp
import pandas as pd
import plotly_express as px
from helpsk.utility import read_pickle, Timer
from helpsk.sklearn_eval import MLExperimentResults

from source.service.model_registry import ModelRegistry

%cd /code

from source.config import config  # noqa: E402
logging.config.fileConfig(
    "source/config/logging_to_file.conf",
    defaults={'logfilename': 'output/log.log'},
    disable_existing_loggers=False,
)

/usr/local/lib/python3.11/site-packages/IPython/core/magics/osm.py:417: UserWarning: using dhist requires you to install the `pickleshare` library.
  self.shell.db['dhist'] = compress_dhist(dhist)[-100:]

/code

Get Latest Experiment Run from MLFlow

registry = ModelRegistry(tracking_uri=config.experiment_server_url())
experiment = registry.get_experiment_by_name(exp_name=config.experiment_name())
logging.info(f"Experiment id: {experiment.last_run.exp_id}")
logging.info(f"Experiment name: {experiment.last_run.exp_name}")
logging.info(f"Run id: {experiment.last_run.run_id}")
logging.info(f"Metric(s): {experiment.last_run.metrics}")

2023-11-24 20:59:39 - INFO     | Experiment id: 1
2023-11-24 20:59:39 - INFO     | Experiment name: credit
2023-11-24 20:59:39 - INFO     | Run id: 7c18722134d54a99991126ac0a1c2971
2023-11-24 20:59:39 - INFO     | Metric(s): {'roc_auc': 0.753377535324465}

Last Run vs Production

What is the metric/performance from the model associated with the last run?

logging.info(f"last run metrics: {experiment.last_run.metrics}")

2023-11-24 20:59:39 - INFO     | last run metrics: {'roc_auc': 0.753377535324465}

What is the metric/performance of the model in production?

production_run = registry.get_production_run(model_name=config.model_name())
logging.info(f"production run metrics: {production_run.metrics}")

2023-11-24 20:59:39 - INFO     | production run metrics: {'roc_auc': 0.753377535324465}

Last Run

# underlying mlflow object
experiment.last_run.mlflow_entity

<Run: data=<RunData: metrics={'roc_auc': 0.753377535324465}, params={'prep__numeric__imputer__transformer': 'SimpleImputer()',
 'prep__numeric__pca__transformer': 'None',
 'prep__numeric__scaler__transformer': 'None',
 'prep__savings_status__savings_encoder__transformer': "OneHotEncoder(handle_unknown='ignore')"}, tags={'mlflow.log-model.history': '[{"run_id": "7c18722134d54a99991126ac0a1c2971", '
                             '"artifact_path": "model", "utc_time_created": '
                             '"2023-11-24 20:59:35.369562", "flavors": '
                             '{"python_function": {"model_path": "model.pkl", '
                             '"predict_fn": "predict", "loader_module": '
                             '"mlflow.sklearn", "python_version": "3.11.6", '
                             '"env": {"conda": "conda.yaml", "virtualenv": '
                             '"python_env.yaml"}}, "sklearn": '
                             '{"pickled_model": "model.pkl", '
                             '"sklearn_version": "1.3.2", '
                             '"serialization_format": "cloudpickle", "code": '
                             'null}}, "model_uuid": '
                             '"8c19cae6bc014da3b8e7ada65bda5e94", '
                             '"mlflow_version": "2.8.0", "model_size_bytes": '
                             '15618339}]',
 'mlflow.note.content': '2023_11_24_20_59_08',
 'mlflow.runName': '2023_11_24_20_59_08',
 'mlflow.source.git.commit': '81a963fcbc4794b8b7bc6c330fc6b034760eb65d',
 'mlflow.source.name': 'source/entrypoints/cli.py',
 'mlflow.source.type': 'LOCAL',
 'mlflow.user': 'root',
 'type': 'BayesSearchCV'}>, info=<RunInfo: artifact_uri='/code/mlflow-artifact-root/1/7c18722134d54a99991126ac0a1c2971/artifacts', end_time=1700859577033, experiment_id='1', lifecycle_stage='active', run_id='7c18722134d54a99991126ac0a1c2971', run_name='2023_11_24_20_59_08', run_uuid='7c18722134d54a99991126ac0a1c2971', start_time=1700859548079, status='FINISHED', user_id='root'>, inputs=<RunInputs: dataset_inputs=[]>>

---

Load Training & Test Data Info

with Timer("Loading training/test datasets"):
    X_train = experiment.last_run.download_artifact(artifact_name='x_train.pkl', read_from=read_pickle)  # noqa
    X_test = experiment.last_run.download_artifact(artifact_name='x_test.pkl', read_from=read_pickle)  # noqa
    y_train = experiment.last_run.download_artifact(artifact_name='y_train.pkl', read_from=read_pickle)  # noqa
    y_test = experiment.last_run.download_artifact(artifact_name='y_test.pkl', read_from=read_pickle)  # noqa

Timer Started: Loading training/test datasets
Timer Finished (0.01 seconds)

logging.info(f"training X shape: {X_train.shape}")
logging.info(f"training y length: {len(y_train)}")

logging.info(f"test X shape: {X_test.shape}")
logging.info(f"test y length: {len(y_test)}")

2023-11-24 20:59:39 - INFO     | training X shape: (800, 20)
2023-11-24 20:59:39 - INFO     | training y length: 800
2023-11-24 20:59:39 - INFO     | test X shape: (200, 20)
2023-11-24 20:59:39 - INFO     | test y length: 200

np.unique(y_train, return_counts=True)

(array([0, 1]), array([559, 241]))

train_y_proportion = np.unique(y_train, return_counts=True)[1] \
    / np.sum(np.unique(y_train, return_counts=True)[1])
logging.info(f"balance of y in training: {train_y_proportion}")

2023-11-24 20:59:39 - INFO     | balance of y in training: [0.69875 0.30125]

test_y_proportion = np.unique(y_test, return_counts=True)[1] \
    / np.sum(np.unique(y_test, return_counts=True)[1])
logging.info(f"balance of y in test: {test_y_proportion}")

2023-11-24 20:59:39 - INFO     | balance of y in test: [0.705 0.295]

Cross Validation Results

Best Scores/Params

results = experiment.last_run.download_artifact(
    artifact_name='experiment.yaml',
    read_from=MLExperimentResults.from_yaml_file,
)
logging.info(f"Best Score: {results.best_score}")
logging.info(f"Best Params: {results.best_params}")

2023-11-24 20:59:39 - INFO     | Best Score: 0.753377535324465
2023-11-24 20:59:39 - INFO     | Best Params: {'model': 'RandomForestClassifier()', 'imputer': 'SimpleImputer()', 'scaler': 'None', 'pca': 'None', 'savings_status_encoder': 'OneHotEncoder()'}

# Best model from each model-type.
data = results.to_formatted_dataframe(return_style=False, include_rank=True)
data["model_rank"] = data.groupby("model")["roc_auc Mean"].rank(method="first", ascending=False)
data.query('model_rank == 1')

	rank	roc_auc Mean	roc_auc 95CI.LO	roc_auc 95CI.HI	model	C	max_features	max_depth	n_estimators	min_samples_split	...	colsample_bytree	colsample_bylevel	reg_alpha	reg_lambda	num_leaves	imputer	scaler	pca	savings_status_encoder	model_rank
10	1	0.753	0.649	0.857	RandomForestClassifier()	NaN	NaN	NaN	NaN	NaN	...	NaN	NaN	NaN	NaN	NaN	SimpleImputer()	None	None	OneHotEncoder()	1.0
5	3	0.750	0.656	0.845	ExtraTreesClassifier()	NaN	NaN	NaN	NaN	NaN	...	NaN	NaN	NaN	NaN	NaN	SimpleImputer()	None	None	OneHotEncoder()	1.0
0	5	0.744	0.641	0.847	LogisticRegression()	NaN	NaN	NaN	NaN	NaN	...	NaN	NaN	NaN	NaN	NaN	SimpleImputer()	StandardScaler()	None	OneHotEncoder()	1.0
19	6	0.744	0.640	0.848	XGBClassifier()	NaN	NaN	1.0	1315.0	NaN	...	0.530125	0.985952	0.001700	2.777691	NaN	SimpleImputer(strategy='median')	None	None	SavingsStatusEncoder()	1.0
23	12	0.732	0.613	0.850	LGBMClassifier()	NaN	NaN	NaN	NaN	NaN	...	0.745611	NaN	0.877292	47.641776	497.0	SimpleImputer(strategy='median')	None	None	OneHotEncoder()	1.0

5 rows × 26 columns

results.to_formatted_dataframe(return_style=True,
                               include_rank=True,
                               num_rows=500)

rank	roc_auc Mean	roc_auc 95CI.LO	roc_auc 95CI.HI	model	C	max_features	max_depth	n_estimators	min_samples_split	min_samples_leaf	max_samples	criterion	learning_rate	min_child_weight	subsample	colsample_bytree	colsample_bylevel	reg_alpha	reg_lambda	num_leaves	imputer	scaler	pca	savings_status_encoder
1	0.753	0.649	0.857	RandomForestClassifier()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	None	None	OneHotEncoder()
2	0.751	0.624	0.878	RandomForestClassifier()	<NA>	0.583	35.000	1,474.000	22.000	5.000	0.765	entropy	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='median')	None	PCA('mle')	OneHotEncoder()
3	0.750	0.656	0.845	ExtraTreesClassifier()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	None	None	OneHotEncoder()
4	0.747	0.631	0.863	RandomForestClassifier()	<NA>	0.239	41.000	1,886.000	3.000	15.000	0.864	gini	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	None	PCA('mle')	SavingsStatusEncoder()
5	0.744	0.641	0.847	LogisticRegression()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	StandardScaler()	None	OneHotEncoder()
6	0.744	0.640	0.848	XGBClassifier()	<NA>	<NA>	1.000	1,315.000	<NA>	<NA>	<NA>	<NA>	0.022	17.000	0.716	0.530	0.986	0.002	2.778	<NA>	SimpleImputer(strategy='median')	None	None	SavingsStatusEncoder()
7	0.743	0.703	0.783	LogisticRegression()	0.000	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	MinMaxScaler()	None	SavingsStatusEncoder()
8	0.739	0.606	0.872	RandomForestClassifier()	<NA>	0.911	74.000	1,265.000	39.000	17.000	0.751	entropy	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	None	PCA('mle')	OneHotEncoder()
9	0.739	0.626	0.851	RandomForestClassifier()	<NA>	0.445	87.000	1,244.000	33.000	27.000	0.795	gini	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	None	PCA('mle')	OneHotEncoder()
10	0.734	0.672	0.796	ExtraTreesClassifier()	<NA>	0.135	15.000	1,987.000	10.000	39.000	0.708	gini	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='most_frequent')	None	None	SavingsStatusEncoder()
11	0.732	0.606	0.859	XGBClassifier()	<NA>	<NA>	20.000	1,733.000	<NA>	<NA>	<NA>	<NA>	0.010	2.000	0.673	0.772	0.881	0.281	2.005	<NA>	SimpleImputer()	None	PCA('mle')	OneHotEncoder()
12	0.732	0.613	0.850	LGBMClassifier()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	0.732	0.746	<NA>	0.877	47.642	497.000	SimpleImputer(strategy='median')	None	None	OneHotEncoder()
13	0.731	0.646	0.816	LGBMClassifier()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	0.630	0.248	<NA>	5.830	44.985	347.000	SimpleImputer(strategy='most_frequent')	None	None	OneHotEncoder()
14	0.727	0.633	0.821	ExtraTreesClassifier()	<NA>	0.502	29.000	1,249.000	25.000	35.000	0.855	gini	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='median')	None	None	OneHotEncoder()
15	0.725	0.590	0.860	LogisticRegression()	0.000	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='most_frequent')	StandardScaler()	PCA('mle')	SavingsStatusEncoder()
16	0.723	0.585	0.862	LogisticRegression()	0.000	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='median')	StandardScaler()	None	SavingsStatusEncoder()
17	0.723	0.588	0.859	LogisticRegression()	0.000	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='median')	StandardScaler()	PCA('mle')	SavingsStatusEncoder()
18	0.723	0.633	0.813	ExtraTreesClassifier()	<NA>	0.537	2.000	1,275.000	14.000	47.000	0.805	entropy	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='median')	None	None	SavingsStatusEncoder()
19	0.722	0.637	0.808	ExtraTreesClassifier()	<NA>	0.768	54.000	909.000	16.000	30.000	0.762	entropy	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer(strategy='median')	None	PCA('mle')	SavingsStatusEncoder()
20	0.721	0.622	0.819	XGBClassifier()	<NA>	<NA>	3.000	1,482.000	<NA>	<NA>	<NA>	<NA>	0.067	18.000	0.889	0.636	0.615	0.000	2.093	<NA>	SimpleImputer(strategy='median')	None	None	OneHotEncoder()
21	0.715	0.667	0.764	LGBMClassifier()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	0.429	0.877	<NA>	13.997	35.364	388.000	SimpleImputer(strategy='most_frequent')	None	None	OneHotEncoder()
22	0.713	0.604	0.823	LGBMClassifier()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	None	None	OneHotEncoder()
23	0.701	0.587	0.814	XGBClassifier()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	None	None	OneHotEncoder()
24	0.691	0.638	0.745	LGBMClassifier()	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	0.965	0.241	<NA>	19.400	3.771	281.000	SimpleImputer(strategy='median')	None	PCA('mle')	OneHotEncoder()
25	0.681	0.582	0.780	XGBClassifier()	<NA>	<NA>	4.000	1,961.000	<NA>	<NA>	<NA>	<NA>	0.271	3.000	0.671	0.797	0.696	0.000	2.150	<NA>	SimpleImputer(strategy='median')	None	PCA('mle')	SavingsStatusEncoder()

results.to_formatted_dataframe(query='model == "RandomForestClassifier()"', include_rank=True)

rank	roc_auc Mean	roc_auc 95CI.LO	roc_auc 95CI.HI	max_features	max_depth	n_estimators	min_samples_split	min_samples_leaf	max_samples	criterion	imputer	pca	savings_status_encoder
1	0.753	0.649	0.857	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	<NA>	SimpleImputer()	None	OneHotEncoder()
2	0.751	0.624	0.878	0.583	35.000	1,474.000	22.000	5.000	0.765	entropy	SimpleImputer(strategy='median')	PCA('mle')	OneHotEncoder()
3	0.747	0.631	0.863	0.239	41.000	1,886.000	3.000	15.000	0.864	gini	SimpleImputer()	PCA('mle')	SavingsStatusEncoder()
4	0.739	0.606	0.872	0.911	74.000	1,265.000	39.000	17.000	0.751	entropy	SimpleImputer()	PCA('mle')	OneHotEncoder()
5	0.739	0.626	0.851	0.445	87.000	1,244.000	33.000	27.000	0.795	gini	SimpleImputer()	PCA('mle')	OneHotEncoder()

results.to_formatted_dataframe(query='model == "LogisticRegression()"', include_rank=True)

rank	roc_auc Mean	roc_auc 95CI.LO	roc_auc 95CI.HI	C	imputer	scaler	pca	savings_status_encoder
1	0.744	0.641	0.847	<NA>	SimpleImputer()	StandardScaler()	None	OneHotEncoder()
2	0.743	0.703	0.783	0.000	SimpleImputer()	MinMaxScaler()	None	SavingsStatusEncoder()
3	0.725	0.590	0.860	0.000	SimpleImputer(strategy='most_frequent')	StandardScaler()	PCA('mle')	SavingsStatusEncoder()
4	0.723	0.585	0.862	0.000	SimpleImputer(strategy='median')	StandardScaler()	None	SavingsStatusEncoder()
5	0.723	0.588	0.859	0.000	SimpleImputer(strategy='median')	StandardScaler()	PCA('mle')	SavingsStatusEncoder()

BayesSearchCV Performance Over Time

results.plot_performance_across_trials(facet_by='model').show()

results.plot_performance_across_trials(query='model == "RandomForestClassifier()"').show()

---

Variable Performance Over Time

results.plot_parameter_values_across_trials(query='model == "RandomForestClassifier()"').show()

---

Scatter Matrix

# results.plot_scatter_matrix(query='model == "RandomForestClassifier()"',
#                             height=1000, width=1000).show()

---

Variable Performance - Numeric

results.plot_performance_numeric_params(query='model == "RandomForestClassifier()"',
                                        height=800)

results.plot_parallel_coordinates(query='model == "RandomForestClassifier()"').show()

---

Variable Performance - Non-Numeric

results.plot_performance_non_numeric_params(query='model == "RandomForestClassifier()"').show()

---

results.plot_score_vs_parameter(
    query='model == "RandomForestClassifier()"',
    parameter='max_features',
    size='max_depth',
    color='savings_status_encoder',
)

---

# results.plot_parameter_vs_parameter(
#     query='model == "XGBClassifier()"',
#     parameter_x='colsample_bytree',
#     parameter_y='learning_rate',
#     size='max_depth'
# )

# results.plot_parameter_vs_parameter(
#     query='model == "XGBClassifier()"',
#     parameter_x='colsample_bytree',
#     parameter_y='learning_rate',
#     size='imputer'
# )

---

Last Run - Test Set Performance

last_model = experiment.last_run.download_artifact(
    artifact_name='model/model.pkl',
    read_from=read_pickle,
)
print(type(last_model.model))

<class 'sklearn.pipeline.Pipeline'>

last_model

SklearnModelWrapper(model=Pipeline(steps=[('prep',
                                           ColumnTransformer(transformers=[('numeric',
                                                                            Pipeline(steps=[('imputer',
                                                                                             TransformerChooser(transformer=SimpleImputer())),
                                                                                            ('scaler',
                                                                                             TransformerChooser()),
                                                                                            ('pca',
                                                                                             TransformerChooser())]),
                                                                            ['duration',
                                                                             'credit_amount',
                                                                             'installment_commitment',
                                                                             'residence_since',
                                                                             'age',
                                                                             'existing_credits',
                                                                             'num_dependents']),
                                                                           ('n...
                                                                             'employment',
                                                                             'personal_status',
                                                                             'other_parties',
                                                                             'property_magnitude',
                                                                             'other_payment_plans',
                                                                             'housing',
                                                                             'job',
                                                                             'own_telephone',
                                                                             'foreign_worker']),
                                                                           ('savings_status',
                                                                            Pipeline(steps=[('savings_encoder',
                                                                                             TransformerChooser(transformer=OneHotEncoder(handle_unknown='ignore')))]),
                                                                            ['savings_status'])])),
                                          ('model',
                                           RandomForestClassifier(n_estimators=500,
                                                                  random_state=42))]))

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

test_predictions = last_model.predict(X_test)
test_predictions[0:10]

array([0.408, 0.522, 0.678, 0.404, 0.088, 0.454, 0.092, 0.492, 0.176,
       0.232])

evaluator = hlp.sklearn_eval.TwoClassEvaluator(
    actual_values=y_test,
    predicted_scores=test_predictions,
    score_threshold=0.37,
)

evaluator.plot_actual_vs_predict_histogram()

evaluator.plot_confusion_matrix()

evaluator.all_metrics_df(return_style=True,
                         dummy_classifier_strategy=['prior', 'constant'],
                         round_by=3)

	Score	Dummy (prior)	Dummy (constant)	Explanation
AUC	0.815	0.500	0.500	Area under the ROC curve (true pos. rate vs false pos. rate); ranges from 0.5 (purely random classifier) to 1.0 (perfect classifier)
True Positive Rate	0.729	0.000	1.000	72.9% of positive instances were correctly identified.; i.e. 43 "Positive Class" labels were correctly identified out of 59 instances; a.k.a Sensitivity/Recall
True Negative Rate	0.801	1.000	0.000	80.1% of negative instances were correctly identified.; i.e. 113 "Negative Class" labels were correctly identified out of 141 instances
False Positive Rate	0.199	0.000	1.000	19.9% of negative instances were incorrectly identified as positive; i.e. 28 "Negative Class" labels were incorrectly identified as "Positive Class", out of 141 instances
False Negative Rate	0.271	1.000	0.000	27.1% of positive instances were incorrectly identified as negative; i.e. 16 "Positive Class" labels were incorrectly identified as "Negative Class", out of 59 instances
Positive Predictive Value	0.606	0.000	0.295	When the model claims an instance is positive, it is correct 60.6% of the time; i.e. out of the 71 times the model predicted "Positive Class", it was correct 43 times; a.k.a precision
Negative Predictive Value	0.876	0.705	0.000	When the model claims an instance is negative, it is correct 87.6% of the time; i.e. out of the 129 times the model predicted "Negative Class", it was correct 113 times
F1 Score	0.662	0.000	0.456	The F1 score can be interpreted as a weighted average of the precision and recall, where an F1 score reaches its best value at 1 and worst score at 0.
Precision/Recall AUC	0.660	0.295	0.295	Precision/Recall AUC is calculated with `average_precision` which summarizes a precision-recall curve as the weighted mean of precisions achieved at each threshold. See sci-kit learn documentation for caveats.
Accuracy	0.780	0.705	0.295	78.0% of instances were correctly identified
Error Rate	0.220	0.295	0.705	22.0% of instances were incorrectly identified
% Positive	0.295	0.295	0.295	29.5% of the data are positive; i.e. out of 200 total observations; 59 are labeled as "Positive Class"
Total Observations	200	200	200	There are 200 total observations; i.e. sample size

evaluator.plot_roc_auc_curve().show()

evaluator.plot_precision_recall_auc_curve().show()

evaluator.plot_threshold_curves(score_threshold_range=(0.1, 0.7)).show()

evaluator.plot_precision_recall_tradeoff(score_threshold_range=(0.1, 0.6)).show()

evaluator.calculate_lift_gain(return_style=True)

/usr/local/lib/python3.11/site-packages/helpsk/sklearn_eval.py:2480: FutureWarning:

The default of observed=False is deprecated and will be changed to True in a future version of pandas. Pass observed=False to retain current behavior or observed=True to adopt the future default and silence this warning.

	Gain	Lift
Percentile
5	0.14	2.71
10	0.22	2.20
15	0.36	2.37
20	0.51	2.54
25	0.54	2.17
30	0.68	2.26
35	0.73	2.08
40	0.76	1.91
45	0.76	1.69
50	0.81	1.63
55	0.85	1.54
60	0.85	1.41
65	0.88	1.36
70	0.90	1.28
75	0.95	1.27
80	0.97	1.21
85	0.98	1.16
90	0.98	1.09
95	1.00	1.05
100	1.00	1.00

---

Production Model - Test Set Performance

production_model = production_run.download_artifact(
    artifact_name='model/model.pkl',
    read_from=read_pickle,
)
print(type(production_model.model))

<class 'sklearn.pipeline.Pipeline'>

production_model

SklearnModelWrapper(model=Pipeline(steps=[('prep',
                                           ColumnTransformer(transformers=[('numeric',
                                                                            Pipeline(steps=[('imputer',
                                                                                             TransformerChooser(transformer=SimpleImputer())),
                                                                                            ('scaler',
                                                                                             TransformerChooser()),
                                                                                            ('pca',
                                                                                             TransformerChooser())]),
                                                                            ['duration',
                                                                             'credit_amount',
                                                                             'installment_commitment',
                                                                             'residence_since',
                                                                             'age',
                                                                             'existing_credits',
                                                                             'num_dependents']),
                                                                           ('n...
                                                                             'employment',
                                                                             'personal_status',
                                                                             'other_parties',
                                                                             'property_magnitude',
                                                                             'other_payment_plans',
                                                                             'housing',
                                                                             'job',
                                                                             'own_telephone',
                                                                             'foreign_worker']),
                                                                           ('savings_status',
                                                                            Pipeline(steps=[('savings_encoder',
                                                                                             TransformerChooser(transformer=OneHotEncoder(handle_unknown='ignore')))]),
                                                                            ['savings_status'])])),
                                          ('model',
                                           RandomForestClassifier(n_estimators=500,
                                                                  random_state=42))]))

In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
On GitHub, the HTML representation is unable to render, please try loading this page with nbviewer.org.

test_predictions = production_model.predict(X_test)
test_predictions[0:10]

array([0.408, 0.522, 0.678, 0.404, 0.088, 0.454, 0.092, 0.492, 0.176,
       0.232])

evaluator = hlp.sklearn_eval.TwoClassEvaluator(
    actual_values=y_test,
    predicted_scores=test_predictions,
    score_threshold=0.37,
)

evaluator.plot_actual_vs_predict_histogram()

evaluator.plot_confusion_matrix()

evaluator.all_metrics_df(return_style=True,
                         dummy_classifier_strategy=['prior', 'constant'],
                         round_by=3)

	Score	Dummy (prior)	Dummy (constant)	Explanation
AUC	0.815	0.500	0.500	Area under the ROC curve (true pos. rate vs false pos. rate); ranges from 0.5 (purely random classifier) to 1.0 (perfect classifier)
True Positive Rate	0.729	0.000	1.000	72.9% of positive instances were correctly identified.; i.e. 43 "Positive Class" labels were correctly identified out of 59 instances; a.k.a Sensitivity/Recall
True Negative Rate	0.801	1.000	0.000	80.1% of negative instances were correctly identified.; i.e. 113 "Negative Class" labels were correctly identified out of 141 instances
False Positive Rate	0.199	0.000	1.000	19.9% of negative instances were incorrectly identified as positive; i.e. 28 "Negative Class" labels were incorrectly identified as "Positive Class", out of 141 instances
False Negative Rate	0.271	1.000	0.000	27.1% of positive instances were incorrectly identified as negative; i.e. 16 "Positive Class" labels were incorrectly identified as "Negative Class", out of 59 instances
Positive Predictive Value	0.606	0.000	0.295	When the model claims an instance is positive, it is correct 60.6% of the time; i.e. out of the 71 times the model predicted "Positive Class", it was correct 43 times; a.k.a precision
Negative Predictive Value	0.876	0.705	0.000	When the model claims an instance is negative, it is correct 87.6% of the time; i.e. out of the 129 times the model predicted "Negative Class", it was correct 113 times
F1 Score	0.662	0.000	0.456	The F1 score can be interpreted as a weighted average of the precision and recall, where an F1 score reaches its best value at 1 and worst score at 0.
Precision/Recall AUC	0.660	0.295	0.295	Precision/Recall AUC is calculated with `average_precision` which summarizes a precision-recall curve as the weighted mean of precisions achieved at each threshold. See sci-kit learn documentation for caveats.
Accuracy	0.780	0.705	0.295	78.0% of instances were correctly identified
Error Rate	0.220	0.295	0.705	22.0% of instances were incorrectly identified
% Positive	0.295	0.295	0.295	29.5% of the data are positive; i.e. out of 200 total observations; 59 are labeled as "Positive Class"
Total Observations	200	200	200	There are 200 total observations; i.e. sample size

evaluator.plot_roc_auc_curve().show()

evaluator.plot_precision_recall_auc_curve().show()

evaluator.plot_threshold_curves(score_threshold_range=(0.1, 0.7)).show()

evaluator.plot_precision_recall_tradeoff(score_threshold_range=(0.1, 0.6)).show()

evaluator.calculate_lift_gain(return_style=True)

/usr/local/lib/python3.11/site-packages/helpsk/sklearn_eval.py:2480: FutureWarning:

The default of observed=False is deprecated and will be changed to True in a future version of pandas. Pass observed=False to retain current behavior or observed=True to adopt the future default and silence this warning.

	Gain	Lift
Percentile
5	0.14	2.71
10	0.22	2.20
15	0.36	2.37
20	0.51	2.54
25	0.54	2.17
30	0.68	2.26
35	0.73	2.08
40	0.76	1.91
45	0.76	1.69
50	0.81	1.63
55	0.85	1.54
60	0.85	1.41
65	0.88	1.36
70	0.90	1.28
75	0.95	1.27
80	0.97	1.21
85	0.98	1.16
90	0.98	1.09
95	1.00	1.05
100	1.00	1.00

---

Feature Importance

try:
    importances = production_model.model['model'].feature_importances_
    feature_names = [
        x.replace('non_numeric__', '').replace('numeric__', '')
        for x in production_model.model[:-1].get_feature_names_out()
    ]
    feature_importances = sorted(
        zip(feature_names, importances, strict=True),
        key=lambda x: x[1],
        reverse=False,
    )
    fig = px.bar(
        pd.DataFrame(feature_importances, columns=['feature', 'importance']).tail(20),
        y='feature',
        x='importance',
        orientation='h',
        height=700,
        width=800,
        title='Feature Importances of Production Model',
    )
    fig.show()
except:  # noqa
    print("Error calculating feature importances.")

---